A Bott-Borel-Weil theory for direct limits of algebraic groups

A Bott-Borel-Weil theory for direct limits of algebraic groups Ivan Dimitrov, Ivan B. Penkov, Joseph Albert Wolf American Journal of Mathematics, Volume 124, Number 5, October 2002, pp. 955-998 (Article) Published by The Johns Hopkins University Press DOI: 10.1353/ajm.2002.0025 For additional information about this article http://muse.jhu.edu/journals/ajm/summary/v124/124.5dimitrov.html Access provided by University of California @ Berkeley (24 Oct 2013 18:56 GMT)

A BOTT-BOREL-WEIL THEORY FOR DIRECT LIMITS OF ALGEBRAIC GROUPS By IVAN DIMITROV, IVAN PENKOV and JOSEPH A. WOLF Abstract. We develop a Bott-Borel-Weil theory for direct limits of algebraic groups. Some of our results apply to locally reductive ind-groups G in general, i.e., to arbitrary direct limits of connected reductive linear algebraic groups. Our most explicit results concern root-reductive ind-groups G, the locally reductive ind-groups whose Lie algebras admit root decomposition. Given a parabolic subgroup P of G and a rational irreducible P-module, we consider the irreducible G-sheaves O G/P (E) and their duals O G/P (E ). These sheaves are locally free, in general of infinite rank. We prove a general analog of the Bott-Borel-Weil Theorem for O G/P (E ), namely that H q (G/P; O G/P (E )) is nonzero for at most one index q = q 0 and that H q 0 (G/P; O G/P (E )) is isomorphic to the dual of a rational irreducible G-module V. For q 0 > 0 we show that (in contrast to the finite dimensional case) V need not admit an irreducible P-submodule. There, however, one has a larger parabolic subgroup w P P, constructed from P and a Weyl group element w of length q 0, such that V is generated by an irreducible w P-submodule. Consequently certain G-modules V can appear only for q 0 > 0, never for q 0 = 0. For O G/P (E) we show that there is no analog of Bott s vanishing theorem, more precisely that O G/P (E) can have arbitrarily many nonzero cohomology groups. Finally, we give an explicit criterion for the projectivity of the ind-variety G/P, showing that G/P is in general not projective. 0. Introduction. The classical Bott-Borel-Weil Theorem [5] is the cornerstone of the geometric approach to representation theory. Analogs of this theorem have been studied in various contexts, including that of homogeneous spaces in characteristic p [14] and of homogeneous superspaces [23]. In these contexts the Bott-Borel-Weil Theorem does not carry over as a single theorem, and this has inspired important areas of investigation. Infinite dimensional group analogs of the Bott-Borel-Weil Theorem have also been studied. The loop group case was addressed in the 1980 s; see [18] and [24]. Direct limit Lie groups were first addressed recently by L. Natarajan, E. Rodríguez-Carrington and one of us [20]. Roughly speaking, the results of [20] extend the finite dimensional analytic Bott-Borel-Weil Theorem to direct limit Lie groups and direct limit unitary representations, in other words to the analytic category with representations on Hilbert spaces. In this paper we view the classical Bott-Borel-Weil Theorem as a result in algebraic geometry, see [6] and [7] (B. Kostant s purely algebraic version of this Manuscript received February 2, 2001; revised August 27, 2001. Research of the first author supported in part by NSF Grant DMS 97-09820 and the Max Planck Institute for Mathematics, Bonn; research of the second author supported in part by NSF Grant DMS 97-09820, the University of Hamburg, and the Max Planck Institute for Mathematics, Bonn; research of the third author supported in part by NSF Grant DMS 97-09820 and 97-05709, and the Alexander von Humboldt Foundation. American Journal of Mathematics 124 (2002), 955 998. 955

956 I. DIMITROV, I. PENKOV, AND J. WOLF theorem, [16], is beyond our scope here), and try to find its analog for direct limits of linear algebraic groups. This turns out to be a very interesting problem and, as in the characteristic p and the supergeometry contexts, one is led not to a single theorem but to the development of a new circle of ideas. To start, recall two forms of the Bott-Borel-Weil Theorem for a connected complex reductive linear algebraic group G: (1) Let B be a Borel subgroup of G, L an irreducible (one dimensional) rational B-module, L G/B the associated homogeneous line bundle, and O G/B (L) be the sheaf of regular local sections of L. Then at most one of the cohomology groups H q (G/B; O G/B (L)) is nonzero, and if H q 0 (G/B; O G/B (L)) = 0 then H q 0 (G/B; O G/B (L)) is an irreducible G-module whose highest weight has an explicit expression in terms of the weight of L. (2) Let P be a parabolic subgroup of G with B P and let π: G/B G/P denote the canonical projection. Let E be an irreducible (finite dimensional) rational P-module, E G/P be the associated homogeneous vector bundle, and O G/P (E) be the sheaf of regular local sections of E. Then O G/P (E) is isomorphic to the direct image sheaf π O G/B (L) for some one dimensional B-module L, and H q (G/P; O G/P (E)) = H q (G/B; O G/B (L)) for any q. Therefore at most one cohomology group H q 0 (G/P; O G/P (E)) is nonzero, and it is an irreducible G- module whose highest weight can be calculated explicitly. Statement (1) admits a reasonably straightforward generalization to locally reductive ind-groups. As follows from Theorem 11.1(ii) below, if G is a locally reductive ind-group, B is a Borel subgroup, and L is any irreducible rational (and thus one dimensional) B-module, then the sheaf O G/B (L) has at most one nonzero cohomology group H q 0 (G/B; O G/B (L)). Furthermore, H q 0 (G/B; O G/B (L)) is isomorphic to the dual of an irreducible rational G-module V, and therefore (in contrast to the classical case and to the analytic category case [20]) is a reducible G-module except in the very special situation of finite dimensional V. Ifq 0 =0, V has a highest weight which is the negative of the weight of L. For q 0 > 0 the explicit description of V in terms of L is an interesting problem, and it is still open for sufficiently general ind-groups G. However, for a root-reductive ind-group G (root-reductive ind-groups are the simplest ind-versions of reductive groups, see Section 4) we prove in Proposition 14.1 that V is a highest weight module whose highest weight is calculated in terms of L using a Weyl group action which has many properties in common with that of the classical Bott-Borel-Weil Theorem. In this paper we develop a Bott-Borel-Weil theory for locally reductive indgroups G, which in particular replaces (2). It involves a number of new constructions, and goes well beyond straightforward generalization. To indicate why this is needed, here are two observations on a parabolic subgroup P of G and an infinite dimensional irreducible rational P-module E. First, E need not be locally irreducible, i.e., need not be isomorphic to a direct limit of irreducible P n -modules, and also E need not be a weight module. In particular, E need not have an extremal weight with respect to any Borel subgroup B P. Second, there

A BOTT-BOREL-WEIL THEORY 957 are two natural G-sheaves to consider: the sheaf O G/P (E) of local sections of the G-bundle E G/P associated to E, and the sheaf O G/P (E ) of local sections of the dual bundle (of uncountable rank) E G/P. It is not difficult to see that the push down approach of (2) applies only to O G/P (E ) and then only under the additional condition that E be a highest weight P-module. Thus the push down approach does not lead to a complete description of the cohomology of O G/P (E) oro G/P (E ). Our main results, collected in Theorem 11.1, concern the sheaf O G/P (E ). When G is root-reductive, but E need not be locally irreducible, we prove that O G/P (E ) has at most one nonzero cohomology group H q 0 (G/P; O G/P (E )), and that this cohomology group is isomorphic to the dual of a rational G-module V. For q 0 = 0, we show that, as in the finite dimensional case, V is parabolically generated by E itself. When E is a weight module this latter result turns out to imply that H 0 (G/P; O G/P (E )) 0 forces E to be locally irreducible; more precisely, E has to be a finite P-module (see Theorem 11.1(iv)). For q 0 > 0, in contrast with the finite dimensional case, V need not be generated by an irreducible P- submodule. We prove that here V is generated by an irreducible submodule of a larger parabolic subgroup w P P that depends on P and a certain element w of the Weyl group. The result is sharp in the sense that there are examples of irreducible P-modules E for which V does not admit an irreducible submodule for any parabolic subgroup of G properly contained in w P; see Example 13.1. This new infinite dimensional phenomenon is quite remarkable, for it provides a geometric construction of G-modules that are not parabolically generated, in particular of cuspidal weight modules [8]. According to Theorem 11.1(iv), cuspidal weight G- modules can only occur in higher cohomology groups. As a consequence, there is no Demazure isomorphism between the unique nonzero higher cohomology group of O G/P (E ) and the zero th cohomology group of O G/P (E ) for any other E. In connection with the above results, we establish an explicit criterion for the projectivity of the ind-variety G/P. Somewhat surprisingly, it turns out that G/P is rarely projective, even for G = GL( ); see 15. Finally, we consider the sheaf O G/P (E) and show by an example that it can have arbitrarily many nonzero cohomology groups. The problem of a systematic description of the cohomology for sheaves of type O G/P (E) remains open. Acknowledgments. We thank Robin Hartshorne, Ziv Ran, and David Ben- Zvi for encouraging discussions at the early stages of this work. In particular, David Ben-Zvi brought [17] to our attention. We thank also Yuri Manin, Vera Serganova and Gregg Zuckerman for constant support throughout the writing of the paper. Notational conventions. The ground field is C, though all results extend easily to any algebraically closed field of characteristic zero. If V is a vector

958 I. DIMITROV, I. PENKOV, AND J. WOLF space, V stands for its dual space. Z + := {0, 1, 2,...}, Z = Z +, N := Z + \{0}, and C is the multiplicative group {a C a = 0}. Expressions such as lim G n and lim H q (X n ;...) denote the direct and inverse limit, respectively of a direct or inverse system indexed by Z +. The notation H q (X; F) always stands for the q th cohomology group of a sheaf F on a topological space X. Ifπ: X Y is a morphism of algebraic varieties, π and π denote respectively the inverse and direct image functors of coherent sheaves. Ind-groups (in particular complex algebraic groups) are denoted by capital letters (e.g. K), and their Lie algebras are denoted by the corresponding lower case Gothic letter (e.g. k). Furthermore, as the ground field is fixed, we will write simply GL(n), SO(n), etc. instead of GL(n, C), SO(n, C), etc. U(k) denotes the universal enveloping algebra of the Lie algebra k. The sign + stands for semidirect sum of Lie algebras: if k = m + n, then m is an ideal in k and n is a complementary subalgebra. The sign f denotes finite or weak direct product of groups or of homogeneous spaces. If G t, t T, is an infinite family of groups, we set f t T G t := { t T g t g t G t for all t and g t = 1 for almost all t T}. f t T (G t /H t ) is the image Similarly, if G t /H t is a family of homogeneous spaces, f of t T G t in t T (G t /H t ). Finally, if k k as a pair of finite dimensional Lie algebras, and E k and E k are respectively a finite dimensional irreducible k- module and a finite dimensional irreducible k -module, we write E k E k if there is an injection of k -modules E k E k. Part I. Ind-varieties and Ind-groups. 1. Ind-varieties. This is a quick summary of the basic definitions on indvarieties. See both [25] and [17] for more detailed summaries. An ind-variety (over C) is a set X with a filtration (1.1) X 0 X 1 X 2 such that X = n 0 X n, each X n is a Noetherian algebraic variety, and the inclusions X n X n+1 are closed immersions of algebraic varieties. In the following we will often write X = lim X n. An ind-variety X is automatically a topological space: a subset U X is open in X if and only if, for each n, U X n is an open subvariety of X n. The sheaf of regular functions on X, orthe structure sheaf O X of X, is the inverse limit O X = lim O Xn of the sheaves of regular functions O Xn on the X n. An ind-variety X = lim X n is proper if and only if all the varieties X n are proper, is affine if and only if all the X n are affine. A morphism from an ind-variety X to an ind-variety Y is a map ϕ: X Y such that, for every

A BOTT-BOREL-WEIL THEORY 959 n 0, the restriction ϕ Xn is a morphism of X n into Y m for some m = m(n). An isomorphism of ind-varieties is a morphism which admits an inverse morphism. An ind-subvariety Z of X is a subset Z X such that Z X n is a subvariety of X n for each n. An ind-variety is connected if it is connected as a topological space. The (Zariski) tangent space T x (X), to an ind-variety X = lim X n at a closed point x X is, by definition, the direct limit lim T x (X n ) where x X n for n sufficiently large. Any ind-variety morphism ϕ: X Y induces linear maps dφ x : T x (X) T φ(x) (Y) for all closed points x X. Example 1.2. (1) C = lim C n = {(a 1, a 2, a 3,...) a n C, only finitely many a n = 0} is an affine ind-variety. (2) Every (complex) vector space V of countable dimension has a canonical structure of an affine ind-variety: any basis {v 1,..., v n,...} identifies V = lim Span{v 1,..., v n } with C as sets and thus defines an ind-variety structure on V. Any other basis of V defines the same ind-variety structure because the identity map on V is an ind-variety isomorphism. (3) Let V be a vector space of countable dimension. Fix an integer k 1. The set Gr(k, V) of all k-dimensional subspaces of V has a canonical structure of proper ind-variety: any filtration 0 V k V k+1 V = r 0 V k+r, dim V k+r = k + r, induces a filtration Gr(k, V k ) Gr(k, V k+1 ) Gr(k, V), and the associated ind-variety structure on Gr(k, V) is independent of choice of filtration on V. For k =1,P(V) :=Gr(1, V) is by definition the projective indspace associated to V. 2. Projective and locally projective ind-varieties. An ind-variety X is locally projective if it admits an ind-variety filtration (1.1) such that all the X n are projective varieties. An ind-variety X is projective if it can be embedded as a closed ind-subvariety into the projective ind-space P(C ). (Tjurin s notion of projectivity [26] is stronger, as it requires the existence of a finite codimensional projective embedding.) Any projective ind-variety is proper and locally projective, but the converse is not true. Below we introduce twisted projective ind-spaces which are the simplest examples of locally projective, generically not projective, ind-varieties. Let Y be a proper Noetherian algebraic variety. A very ample invertible O Y - module L Y (i.e., a locally free sheaf L Y of O Y -modules of rank 1 which is generated by its global sections) determines a closed immersion of Y into the projective space P := P(H 0 (Y; L Y ) ), and L Y is identified with the inverse image under this immersion of the standard sheaf O P (1) on P. See [12, II, 5]. If X is a proper ind-variety, consider an invertible O X -module L X and an ind-variety

960 I. DIMITROV, I. PENKOV, AND J. WOLF filtration (1.1) such that all inverse images L Xn of L X are very ample and the restriction maps H 0 (X n ; L Xn ) H 0 (X n 1 ; L Xn 1 ) are surjective. Then the system dual to the inverse system H 0 (X n ; L Xn ) H 0 (X n 1 ; L Xn 1 ) H 0 (X 0 ; L X0 ) 0 is a direct system of injections, and L X defines a closed immersion of X into P( lim H 0 (X n ; L Xn ) ). We define X to be a twisted projective ind-space if it admits a filtration (1.1) such that X n is a projective space for all n. To a twisted projective ind-space we attach its twisting sequence {c 1, c 2,...}, where c n is the natural number that denotes the first Chern class of the inverse image of O Xn+1 (1) on X n. Any sequence of natural numbers can be obtained as a twisting sequence. PROPOSITION 2.1. A twisted projective ind-space X is projective if and only if its twisting sequence stably equals the sequence {1, 1,...}, i.e. c n =1for n sufficiently large. Proof. If c n = 1 for all n > n 0, then X n0 X n0 +1 is an indvariety filtration for X and, for any n n 0, the invertible O Xn -module O Xn (1) with Chern class 1 is the inverse image of a well-defined O X -module O X (1). Then O X (1) establishes an isomorphism between X and the projective ind-space P( lim H 0 (X n ; O Xn (1)) ). Thus X itself is isomorphic to a projective ind-space. Conversely, let the sequence c 1, c 2,... corresponding to X admit a subsequence c n1, c n2,... with c nt 2 for all t. Assume that X is a closed ind-subvariety of a projective ind-space P and consider the inverse images O P (1) X n of O P (1) on X n, where O P (1) := lim O Pn (1). Denote the Chern class of O P (1) X n by C n. Then, for any s > k, C k /C s = c k+1 c k+2 c s. Therefore the positive integer C n0 is divisible by all products c n0 +1c n0 +2 c n for any n > n 0, which is an obvious contradiction. We remark that the ind-grassmannian Gr(k, V) of Example 1.2(3) is projective, and that the classical Plücker embeddings induce a closed immersion Pl k : Gr(k, V) P( k V) where k denotes k th exterior power. 3. Ind-groups and direct limit Lie algebras. An algebraic ind-group, or, briefly, ind-group, is an ind-variety G with group structure such that the map G G G, (g 1, g 2 ) g 1 g 1 2 is a morphism of ind-varieties. By definition, an ind-group homomorphism is a group homomorphism κ: G K that is also an ind-variety morphism. An ind-subgroup K of G is a subgroup K G that is an ind-subvariety.

A BOTT-BOREL-WEIL THEORY 961 We define a locally linear ind-group as an (affine) ind-variety G with an ind-variety filtration (3.1) G 0 G 1 G 2 such that all the G n are linear algebraic groups and all the inclusions are closed immersions that are group morphisms. Clearly, every locally linear ind-group is an affine ind-group but the converse is not true. An example of an affine indgroup which is not locally linear is provided by the automorphism group Aut(A m ) of the m-dimensional affine space A m ; see [25]. We will only study connected locally linear ind-groups, and all ind-groups considered below are assumed to be connected and locally linear. By a slight abuse of language we refer to them simply as ind-groups. Let G be a (connected) ind-group. In this paper we define a parabolic subgroup of G as an ind-subgroup P of G such that, for a suitable filtration (3.1), P n := P G n is a parabolic subgroup of G n for each n, and, in addition, U Pn 1 = U Pn P n 1, where U Pi denotes the unipotent radical of P i. Similarly, a Borel subgroup of G is an ind-subgroup B of G such that, for a suitable filtration (3.1), B n := B G n is a Borel subgroup of G n for each n (the condition on the unipotent radicals is automatic here), and a Cartan subgroup of G is an ind-subgroup H of G such that, for a suitable filtration (3.1), H n := H G n is a Cartan subgroup of G n for each n. An element u G is unipotent if for some (or equivalently, for any) ind-group filtration (3.1) u is unipotent in G n whenever n is large enough so that u G n. The unipotent radical U G of G is defined to be the largest closed normal ind-subgroup of G such that every element u is unipotent. Finally, an ind-group G is locally reductive if we can choose the ind-group filtration (3.1) so that each G n is a reductive linear algebraic group. Whenever G = lim G n is locally reductive we will assume that the G n are also reductive. Throughout the rest of the paper we will consider ind-groups G with a fixed filtration (3.1) of connected linear algebraic groups. We will assume without explicit mention that the parabolic, Borel and Cartan subgroups of G are aligned with respect to that filtration in the above sense. PROPOSITION 3.2. Let G = lim G n be a locally reductive ind-group and let P = lim P n be a parabolic ind-subgroup. (i) The unipotent radical U P of P is well-defined, and U P = lim U Pn. (ii) The Chevalley semidirect product decompositions P n = U Pn P red n, into the unipotent radical and a complementary reductive subgroup, can be chosen so that P red n P red n+1 for all n. Then P = U P P red where P red := lim P red n. Proof. According to the definition, lim U Pn is a well-defined ind-subgroup of G, which is closed and normal in P since, for each n, U Pn is closed and normal in P n. Furthermore, lim U Pn is the largest closed normal subgroup of P in which every element is unipotent, because the existence of a larger such subgroup would

962 I. DIMITROV, I. PENKOV, AND J. WOLF contradict the fact that U Pn is the unipotent radical of P n for every n. Therefore lim U P n is the unipotent radical U P of P. A theorem of Mostow [19, Theorem 7.1] ensures that the maximal reductive subgroups of P n+1 are just the reductive subgroups R such that P n+1 = U Pn+1 R, and that any two such groups R are conjugate by an element of U Pn+1. See [13, VIII, Theorem 4.3] for a systematic development. Therefore, given a maximal reductive subgroup P red n in P n, we can choose P red n+1 to be any maximal reductive subgroup of P n+1 that contains it. This implies (ii), and the equality P = U P P red follows. The Lie algebra of an ind-group G = lim G n is the direct limit Lie algebra g = lim g n for the direct system (3.3) g 0 g 1 g 2, where g n is the Lie algebra of G n and the inclusions g n g n+1 are the differentials of the group immersions G n G n+1. This Lie algebra is the tangent space T 1G (G) with its natural Lie algebra structure [17]. An ind-group homomorphism κ: G K induces a Lie algebra homomorphism dκ: g k [17]. We shall consider direct limit Lie algebras more generally, sometimes without regard to ind-groups. These will always correspond to direct systems of injections of finite dimensional Lie algebras. Let g be a direct limit Lie algebra. In this paper we define a subalgebra p g to be a parabolic subalgebra if, for a suitable direct system {g n }, p n := p g n is a parabolic subalgebra of g n for each n, and, in addition, u pn 1 = u pn p n 1, where u pi denotes the nilpotent radical of p i. Similarly, a subalgebra b g (respectively h g) isaborel (respectively Cartan) subalgebra of g if, for a suitable direct system {g n }, each b n := b g n (respectively h n := h g n ) is a Borel subalgebra (respectively Cartan subalgebra) of g n. (This definition of parabolic and Borel subalgebras is more general than that of [8].) In the rest of the paper we will automatically assume that g is equipped with a fixed filtration (3.3) and that all parabolic, Borel or Cartan subalgebras we consider are aligned with respect to that filtration in the above sense. We define a direct limit Lie algebra g to be locally reductive if we have an expression g = lim g n where each of the g n is reductive. Whenever we express a locally reductive Lie algebra g as lim g n, it will be assumed that the g n are reductive. The following Proposition is the Lie algebra analog of Proposition 3.2. Part (ii) of this Proposition is an adaptation of a result of Baranov [2, Lemma 4.3]. We leave the proof to the reader. PROPOSITION 3.4. Let g = lim g n be a locally reductive direct limit Lie algebra and let p be a parabolic subalgebra. (i) u p := lim u pn is a well-defined ideal in p. By definition, u p is the nilpotent radical of p.

A BOTT-BOREL-WEIL THEORY 963 (ii) One can choose semidirect sum decompositions p n = u pn + p red n, into the nilpotent radical of p n and a complementary reductive subalgebra, such that each p red n p red n+1. Then p = u p + p red where p red := lim p red n. (iii) Let g be the Lie algebra of a reductive ind-group G, P is a parabolic indsubgroup of G with Lie algebra p, and U P be the unipotent radical of P. Then u p is the Lie algebra of U P. Furthermore, if we choose p red n to be the Lie algebra of P red n of Proposition 3.2 (ii), then p red is the Lie algebra of P red. 4. Root-reductive ind-groups and parabolic subgroups. Let g and g be Lie algebras with root decomposition, so g = h ( α (g ) α ) and g = h ( α (g ) α ). Here h and h are respectively Cartan subalgebras, and and are the corresponding root systems of g and g. A Lie algebra homomorphism ϕ: g g is a root homomorphism, ifϕ(h ) h and ϕ maps every root space (g ) α into a root space (g ) α, thus mapping into. To be precise one should write ϕ: (g, h, ) (g, h, ), but we often leave this to be understood by the reader. A root subalgebra of g is the image of a root homomorphism. A locally reductive direct limit Lie algebra g is root-reductive if it can be expressed as a direct limit g = lim g n where all g n are finite dimensional and reductive, h = lim h n is a Cartan subalgebra, and each injection g n g n+1 carries h n into h n+1 and is a root homomorphism. Finally, a locally reductive ind-group G is root-reductive if its Lie algebra g is root-reductive. Let g = lim g n be a root-reductive direct limit Lie algebra, expressed as a direct limit of root injections (g n, h n, n ) (g n+1, h n+1, n+1 ). Then g admits a root decomposition with respect to the Cartan subalgebra h = n h n. The root system of (g, h) is = n n, and all root spaces g α are one dimensional. It is easy to check (see [8]) that the direct limit W = lim W n of the Weyl groups W n (considered as the groups generated by root reflections) of g n is well defined. In this paper W is by definition the Weyl group of g. Let furthermore b g be a Borel subalgebra such that h b. It determines a decomposition = + such that = +, the positive roots + being the roots of b. We say that a positive root α is b-simple if α cannot be decomposed as the sum of two positive roots. A Weyl group element w W is of finite length with respect to b if w is a (finite) product of simple root reflections, w = σ α1 σ αk for some b-simple roots α 1,..., α k. The length of the shortest such expression is by definition the length of w with respect to b. We remark also that a subalgebra p of a rootreductive direct limit Lie algebra g is parabolic if and only if p n is a parabolic subalgebra of g n for each n, as in this case the condition on the nilpotent radicals is automatically satisfied. Root-reductive direct limit Lie algebras were introduced and studied in [8]; also see [20, Section 7]. They are closely related to the classical simple direct limit Lie algebras a( ), b( ), c( ) and d( ) defined by letting g n be the corresponding finite dimensional Lie algebra of rank n and by requiring that all ϕ n be root injections. The isomorphism class of the resulting direct limit Lie

964 I. DIMITROV, I. PENKOV, AND J. WOLF algebra does not depend on the injections ϕ n [8]. Moreover, every simple infinite dimensional root-reductive Lie algebra is isomorphic to one of the four classical simple direct limit Lie algebras. (A further interesting fact is that b( ) and d( ) are isomorphic as Lie algebras, see [3] and [22]. However, no such isomorphism is a root isomorphism.) We give now an explicit description of the classical simple ind-groups A( ), B( ), C( ), and D( ) whose Lie algebras are respectively a( ), b( ), c( ), and d( ). A( ). Here G = A( ) = lim A(n 1) where G n = A(n 1) = SL(n) and the ( ) inclusion A(n 1) A(n) is given by g. Fixing the Cartan subalgebra g 0 0 1 h of all diagonal matrices in g = a( ), we have = {ε i ε j i = j} where ε i h is given by ε i ( diag{t 1, t 2,...}) =t i. The Weyl group W consists of all permutations of {ε i } which leave all but finitely many ε i fixed. B( ). Here G = B( ) = lim B(n), where B(n) =SO(2n + 1) is the complex special orthogonal group corresponding to the nondegenerate symmetric bilinear form (u, v) =u 1 v 1 + n 1 (u 2i v 2i+1 + u 2i+1 v 2i )onc 2n+1, and where the inclusion B(n) B(n + 1) is given by g ( g 0 0 0 1 0 0 0 1 ). Fixing the Cartan subalgebra h of all diagonal matrices in g = b( ), we have = {±ε i, ±ε i ± ε j i = j} where ε i h is given by ε i ( diag{0, t 1, t 1, t 2, t 2,...}) =t i. The Weyl group W consists of all signed permutations of {ε i } which leave all but finitely many ε i fixed. C( ). Here G = C( ) = lim C(n), where C(n) =Sp(2n) is the complex symplectic group corresponding to the nondegenerate antisymmetric bilinear form u, v = n 1 (u 2i 1 v 2i u 2i v 2i 1 )onc 2n, and where the inclusion C(n) C(n +1) is given by g ( g 0 0 0 1 0 0 0 1 ). Fixing the Cartan subalgebra h of all diagonal matrices in g = c( ), we have = {±2ε i, ±ε i ±ε j i = j} where ε i h is given by ε i ( diag{t 1, t 1, t 2, t 2,...}) =t i. The Weyl group W consists of all signed permutations of {ε i } which leave all but finitely many ε i fixed. D( ). Here G = D( ) = lim D(n), where D(n) =SO(2n) is the complex special orthogonal group corresponding to the nondegenerate symmetric bilinear form (u, v) = n 1 (u 2i 1 v 2i + u 2i v 2i 1 )onc 2n, and where the inclusion D(n) D(n + 1) is given by g ( g 0 0 0 1 0 0 0 1 ). Fixing the Cartan subalgebra h of all diagonal matrices in g = d( ), we have = {±ε i ± ε j i = j} where ε i h is given by ε i ( diag{t 1, t 1, t 2, t 2,...}) =t i. The Weyl group W consists of all signed permutations of {ε i } which leave all but finitely many ε i fixed.

A BOTT-BOREL-WEIL THEORY 965 If g is a root-reductive direct limit Lie algebra and k is a root subalgebra, we set ss k := k ( k ) (where k denotes the set of roots of k) and define k ss to be the Lie subalgebra of k generated by α ss kα. k The Lie algebra part of the following Proposition 4.1 reformulates [8, Theorem 1]. It describes the relationship between an arbitrary root-reductive direct limit Lie algebra and the classical simple direct limit Lie algebras. The group level statements follow from the algebra level statements. PROPOSITION 4.1. Let G be a root-reductive ind-group and g be its Lie algebra. (i) g = g ss + a for some abelian Lie subalgebra a g. Furthermore, h = h ss a where h ss := h g ss. (ii) G has a connected closed normal ind-subgroup G ss with Lie algebra g ss and a connected abelian ind-subgroup A with Lie algebra a, and (g, a) ga defines a homomorphism of the semidirect product G ss A onto G with discrete kernel Z. Furthermore H = (H ss A)/Z where H ss := H G ss has Lie algebra h ss. (iii) g ss = t T gt where the g t are classical simple direct limit algebras or simple finite dimensional Lie algebras. G ss = ( f t T G t )/Z 1 where G t is the connected ind-group of G with Lie algebra g t and where Z 1 is a discrete central subgroup of G ss. Example 4.2. Set GL( ) := limgl(n) where the inclusion GL(n) GL(n+1) ( ) is given by g. Denote A := {diag {a,1,1,1,...} a C } = C. g 0 0 1 Then GL( ) = SL( ) A under g (gα 1, α) where α := diag {det (g), 1, 1, 1,...}. Throughout this paper, when considering a root-reductive Lie algebra g we fix a Cartan subalgebra h = n h n corresponding to a fixed system of root injections g n g n+1. Without loss of generality we assume that any Borel or parabolic subalgebras of g are chosen so that they contain h. Parabolic subgroups (including Borel subgroups) of a root-reductive ind-group G will thus contain the Cartan subgroup H = n H n with Lie algebra h. When we refer to A( ), B( ), C( ), D( ) or GL( ) (or to their Lie algebras) we will furthermore assume that the G n (or g n ), as well as the Cartan subgroup H (respectively, the Cartan subalgebra h) are chosen precisely as in our above explicit description. We now define the parabolic subgroup w P needed in the statement of Theorem 11.1 below. If P is a proper parabolic subgroup (containing H) of a rootreductive ind-group G and p is the Lie algebra of P, then h+p ss is a natural choice for p red. Let w W be a Weyl group element. We define w p to be the parabolic subalgebra of g generated by h and the h-root spaces g α for α p w( p red). Then w P is the parabolic subgroup of G with Lie algebra w p. The subgroup w P is not necessarily proper, as shown by the following example.

966 I. DIMITROV, I. PENKOV, AND J. WOLF Example 4.3. Let G = GL( ), let p gl( ) be the (maximal) parabolic subalgebra with roots {ε 1 ε i 2 i} {ε i ε j 2 i = j}, and let P G be the corresponding parabolic subgroup. If w = σ α for α = ε 1 ε 2, then w P = G. Next we describe the parabolic and Borel subalgebras of root-reductive direct limit Lie algebras g. In view of Proposition 4.1 it suffices to describe parabolic and Borel subalgebras for classical simple g. This gives also a description of the parabolic and Borel subgroups of root-reductive ind-groups G, for if P is a parabolic subgroup of G with Lie algebra p, then P = {g G Ad(g)p = p}. The following statement, which of course is standard in the finite dimensional case, is a variation of Proposition 5 from [8]. PROPOSITION 4.4. Suppose either that g is a classical simple direct limit Lie algebra or that g = gl( ). Let p g be a parabolic subalgebra. Then, for some index set S p, there is an isomorphism of Lie algebras (4.5) p red = s S p l s, where each l s is a subalgebra of p red isomorphic to gl(n), togl( ), to a finite dimensional simple Lie algebra, or to a classical simple direct limit Lie algebra. Furthermore, if g is classical simple S p can be chosen so that l s is isomorphic to gl(n) or gl( ) for any s S p except for at most one index s 0 S p for which l s 0 is isomorphic to: a(n) ora( ) for g = a( ); a(n), a( ), c(n) orc( ) for g = c( ); a(n), a( ), b(n) orb( ) for g = b( ); a(n), a( ), d(n) ord( ) for g = d( ). For a classical simple g, there is a natural choice for the set S p, and moreover S p is linearly ordered. Indeed, fix a Borel subalgebra b of p. For every n the parabolic subalgebra p n = g n p determines marked nodes in the Dynkin diagram of g n which correspond to simple roots α such that both α and α are roots of p n. Let S pn be the set whose elements are all the unmarked nodes together with all connected components of marked nodes. We fix an order on the nodes of the Dynkin diagram of g n which is increasing from left to right. For g = d( ) this order is unique. For g = d( ), if the two rightmost nodes are both marked or both unmarked, we set the upper one to precede the lower one, otherwise, we set the marked one to precede the unmarked one. This order on the Dynkin diagram induces an order on S pn, and it is straightforward to check that the orders on S pn are compatible for different n. Hence they determine a linear order on the union n S p n. In what follows, S p will be fixed as the union n S p n. Note that the order on S p depends only on p and not on b. Furthermore, if s 0 S p and l s 0 is not isomorphic to gl( ), gl(n), a( ) ora(n), then s 0 is necessarily the unique maximal element of S p.

A BOTT-BOREL-WEIL THEORY 967 5. Diagonal ind-groups and beyond. The class of root-reductive indgroups is part of the more general class of diagonal ind-groups. The corresponding class of Lie algebras has been studied quite extensively; see [2], [4], [1] and [20]. An essential difference between root-reductive direct limit Lie algebras and general diagonal direct limit Lie algebras is that the latter need not admit Cartan subalgebras that yield a root decomposition compatible with the direct limit. In this paper we do not develop a complete Bott-Borel-Weil theory for diagonal ind-groups. Here is a diagonal ind-group that is not root-reductive. Consider the sequence of closed immersions GL(2 n ) GL(2 n+1 ), ( g g 0 0 g and let GL(2 ) denote the corresponding ind-group. Its Lie algebra gl(2 )isa diagonal direct limit Lie algebra [1]. If H = lim H n where H n denotes the diagonal matrices in GL(2 n ), then H is a Cartan subgroup of GL(2 ) and gl(2 ) has no root decomposition with respect to h. Let n = {ε i,n ε j,n 1 i, j 2 n } be the root system of gl(2 n ). The Borel subalgebras of gl(2 ) containing h are in bijective correspondence with the systems of triangular decompositions n = + n n satisfying the following compatibility condition: if ε i,n ε j,n + n, then ε i,n+1 ε j,n+1 + n+1 and ε 2 n +i,n+1 ε 2 n +j,n+1 + n+1. One can give a similar description of all parabolic subalgebras containing h in terms of compatible systems of partitions n = + n 0 n n. Here is another interesting ind-group. Its Lie algebra was introduced in [1]. Fix k N, k > 1. Let the κ n : PGL(k 2n ) PGL(k 2n+1 ) be the unique closed immersions of algebraic groups for which the diagrams GL(k 2n ) Ad GL(k 2n+1 ) pr n pr n+1 PGL(k 2n ) κ n PGL(k 2n+1 ) are commutative, where pr n and pr n+1 are the natural projections. We define the reductive ind-group PGL Ad (k 2 ) := lim PGL(k 2n ). If H n PGL(k 2n ) denotes the subgroup of diagonal matrices, then κ n (H n ) H n+1 and H := lim H n is a Cartan subgroup of PGL Ad (k 2 ). A Borel subgroup B PGL Ad (k 2 ) is determined by the B n := B PGL(k 2n ). Given B n, we describe the Borel subgroups of PGL(k 2n+1 ) that contain H n+1 and κ n (B n ), providing recursive descriptions of all Borel subgroups of PGL Ad (k 2 ) that contain H. Note first that specification of B n+1 is the same as specification of a prn+1 1 (H n+1)-invariant maximal flag in the natural representation space of GL(k 2n+1 ), and that natural representation space is the adjoint representation space for GL(k 2n ). Hence, a prn+1 1 (H n+1)-invariant maximal flag in the natural ),

968 I. DIMITROV, I. PENKOV, AND J. WOLF representation of GL(k 2n+1 ) is determined by a linear order on root basis of gl(k 2n ), i.e. a basis (5.1) {x α } α n {h i } 1 i k 2n consisting of root vectors x α and of a basis {h i } 1 i k 2n of h n, where n denotes the root system of gl(k 2n ). The Borel subgroup B n determines a partition n = + n n and thus also the following partial order > on n : for α, β + n, α>βwhenever the b n -height of α is greater then the b n -height of β; α>βfor any α + n and any β n ; and finally, for α, β n, α > β whenever β > α. Now, since n is naturally identified with the set {x α } α n, we can consider all extensions of this partial order on n to a linear order > on the root basis (5.1) such that x α > h i > x β for all i whenever α + n and β n. Any such extension determines a unique prn+1 1 (H n+1)-invariant maximal flag in the natural representation and thus a unique Borel subgroup of PGL(k 2n+1 ). One can check that the Borel subgroups obtained in this way are precisely the Borel subgroups of PGL(k 2n+1 ) that contain H n+1 and κ n (B n ). Part II. Representations. 6. Rational and pro-rational G-modules. Let G = lim G n be an ind-group. We define a G-module to be a vector space V endowed with G n -module structures ϕ n : G n V V (C-linear in V) such that ϕ n+1 Gn V = ϕ n for all n. The maps ϕ n define the structure map ϕ: G V V of the G-module V. We say that V is a rational G-module if in addition the dimension of V is countable and ϕ is a morphism of ind-varieties, where V has the canonical ind-variety structure of Example 1.2. An equivalent definition of a rational G-module: V is isomorphic to the limit of a direct system of injections of rational finite dimensional G n - modules ϕ n : G n V n V n. Every rational G-module is locally finite, in other words, ϕ(g n Cv) generates a finite dimensional submodule for every v V and every n. Any rational module over a reductive algebraic group is completely reducible. Thus, if G is locally reductive and V = lim V n is a rational G-module, then V is a completely reducible G n -module for every n. The category of rational G-modules is too restrictive for our purposes. For example the dual of a rational G-module is no longer rational (as dim V is in general uncountable). We define a G-module U to be pro-rational if it is the dual of a rational G-module. This is equivalent to saying that U is isomorphic to the inverse (or projective) limit lim U n of a system ψn+1 U n ψ n U n 1 ψ n 1 ψ 1 U0 0

A BOTT-BOREL-WEIL THEORY 969 of finite dimensional rational G n -modules U n, where ψ n is a rational G n 1 -module surjection for each n. The G-module structure on lim U n :ifu =(..., u k, u k 1,..., u 0 ) lim U n, so each u n U n and ψ n (u n )=u n 1 for n 1, and if g m G m, then g m u := (..., g m u k, g m u k 1,..., g m u m, ψ m (g m u m ),..., ψ 1 ψ m (g m u m )) where k > m. 7. g-modules. If G is an ind-group and V is a rational G-module with structure map ϕ: G V V, then V is a module for the Lie algebra g of G with structure map dϕ: g V V. We say that a g-module V is rationally G-integrable (or, briefly, G-integrable) if it is obtained by this construction from a rational G-module structure on V. AsG is assumed to be connected, that G- module structure is unique. Furthermore, a G-integrable g-module V necessarily is locally finite. By definition this means that V is locally finite as g n -module for every n, in other words, dim U(g n ) v < for every v V and every n. (In some works on Lie algebra representations, in particular in [15] and [8], the term integrable is a synonym for various versions of local finiteness. This is not acceptable in the present paper because such representations need not integrate from the Lie algebra to the ind-group.) It is straightforward to verify that a countable-dimensional g-module V is locally finite if and only if V is isomorphic to a direct limit lim V n of finite dimensional g n -modules V n. Unless the contrary is stated explicitly, in what follows we will assume automatically that an expression of a locally finite g-module as lim V n corresponds to a direct system of injections V n V n+1. In the rest of this section, G is a locally reductive ind-group, g is its Lie algebra, and p is a parabolic subalgebra of g. We study irreducible locally finite p-modules and parabolically generated irreducible g-modules. PROPOSITION 7.1. Let p be a parabolic subalgebra of g. Let E be an irreducible locally finite p-module and let u p denote the nilpotent radical of p. Then u p E =0. Proof. Suppose that ξ e = 0 for some ξ u p and e E. AsE is irreducible there exists u U(p) such that u ξ e = e. Let n be sufficiently large so that ξ, u U(p n ). Then E n := u pn E n is a proper p n -submodule of E n := U(p n ) e such that e / E n. But as ξ e E n, the equality u ξ e = e is contradictory. Therefore ξ e cannot be nonzero. COROLLARY 7.2. If P is a parabolic subgroup of G and E is an irreducible rational P-module, then the unipotent radical U P of P acts trivially on E. Let E be an irreducible p-module as above. We introduce the induced g- module Ṽ(E) :=U(g) U(p) E.

970 I. DIMITROV, I. PENKOV, AND J. WOLF If p is a Borel subalgebra then dim E = 1, i.e., E = C µ for some µ h, and Ṽ(C µ ) is called a Verma module. In general, dim E =, but nevertheless the following proposition holds. It is very similar to Theorem 2.1 in [1] and therefore we omit the proof. PROPOSITION 7.3. Let E be an irreducible locally finite p-module. Then Ṽ(E) has a unique maximal proper g-submodule I E. An important corollary of Proposition 7.3 is that Ṽ(E) has a unique irreducible quotient module V(E) :=Ṽ(E)/I E. We call V(E) the irreducible g-module parabolically generated by E. Any g-module V which is generated by a p-submodule isomorphic to E has the g-module V(E) as a quotient. Indeed, V admits an obvious surjection σ: Ṽ(E) V whose kernel is necessarily a g-submodule of I E. Thus σ induces a g-surjection s: V =(Ṽ(E)/ ker σ) V(E) =Ṽ(E)/I E. In general, V(E) is not locally finite, and therefore is not rationally G- integrable. The problem of characterizing, for a fixed p, all irreducible locally finite p-modules E for which V(E) isg-integrable, is a generalization of the problem of computing all dominant integral weights for a finite dimensional reductive group. The following Proposition reduces this problem in an explicit way to the structure of E. The case when p is a Borel subalgebra of a classical simple linear Lie algebra was studied in [1] and [21]. PROPOSITION 7.4. Let E be an irreducible locally finite p-module. Then V(E) is G-integrable if and only if E is P-integrable and, for any expression E = lim E n, V(E n ) is a finite dimensional g n -module for each n. When these conditions hold, each V(E n ) is a well-defined G n -module and we have a canonical G-module isomorphism V(E) = lim V(E n ). Proof. First, if E is P-integrable and V(E n ) is finite dimensional for each n, the standard theory of connected algebraic groups applied to G n implies that V(E n ) is G n -integrable for any n. Furthermore, lim V(E n ) is a rational G-module which admits a P-module injection i: E lim V(E n ) such that i(e) generates lim V(E n ) as a G-module. This is sufficient to conclude that lim V(E n )=V(E). Indeed, one need only check that the g-surjection s: lim V(E n ) V(E) induced by i is an isomorphism. Assuming that ker s = 0 we find an n such that ker s V(E n ) = 0. Then, as V(E n ) is an irreducible g n -module, V(E n ) ker s. Therefore ker s i(e n ) = 0, which contradicts the injectivity of i. Conversely, if V(E) isg-integrable, E must be P-integrable. Indeed, if E = lim E n, then for each n E n is a p n -submodule of the finite dimensional G n - submodule of V(E) generated by E n. Therefore, again the theory of connected algebraic groups implies that E n is necessarily P n -integrable. Thus lim E n is a P-integrable p-module. To complete the proof we need to show also that V(E n ) is finite dimensional for each n (and is thus a G n -module) whenever V(E) is G-integrable. This follows from a standard geometric version of Frobenius

A BOTT-BOREL-WEIL THEORY 971 Reciprocity and we present this argument in Section 12 in the proof of Theorem 11.1(i). In the following, we say that an irreducible rational P-module E is dominant if V(E) is a locally finite g-module, and thus is a well-defined irreducible rational G-module. An integral weight λ h is B-dominant for a Borel subgroup B G if the one dimensional B-module C λ of weight λ is dominant. We conclude this Section by recalling some basic definitions for weight modules. Let g be root-reductive. A g-module V is a weight module if it has an h-module decomposition (7.5) V = µ h Vµ, where h is the fixed Cartan subalgebra of g and V µ := {v V h v = µ(h)v for any µ h }. The support supp V of a weight module V is {µ h V µ = 0}. A weight g-module V is finite if the support of V is finite in the direction of every root of g, i.e., if for every µ supp V and every α, the intersection {µ+kα k Z + } supp V is finite. See [8]. A finite g-module is locally finite. In the following we will consider finite g- and p-modules, the latter being defined as p-modules which are weight modules (i.e., which satisfy (7.5)) and which are finite as p red -modules. Also, we define a rational G- orp-module to be finite if it is finite respectively as a weight g- or p-module. 8. Irreducible rational G-modules. Let G be a reductive ind-group. An irreducible rational G- (or P-) module V is called locally irreducible if V = lim V n for some direct system of irreducible rational (finite dimensional) G n - (or P n -) modules V n. Our starting point in this section is that an irreducible rational G- module V is not necessarily locally irreducible. Here are some examples. Example 8.1. For every n fix a pair of nonisomorphic irreducible finite dimensional G n -modules U n, W n such that U n, W n U n+1, and U n, W n W n+1, where here the sign indicates the existence of a G n -module injection. Extend the diagonal injection U n U n U n to a G n -module injection ηn U : U n U n+1 W n+1. Similarly, fix a G n -module injection ηn W : W n U n+1 W n+1. Define η n : U n W n U n+1 W n+1 by η n := ηn U ηn W. Let V n := U n W n. Then the G-module V = lim V n is an irreducible rational G-module that is not locally irreducible. It is irreducible as a consequence of Proposition 8.3 below; see Example 8.4(1). It is not locally irreducible because, for every nonzero v V and for sufficiently large n, the G n -submodule of V generated by v is isomorphic to U n W n, and thus is reducible. Here is an explicit choice of the G n -modules U n and W n. Let G = GL( ). Let U n and W n be the respective irreducible GL(n)-modules with highest weights λ n := ε 1 (n 1)ε n and µ n := ε 1 nε n. The standard branching rule [9] ensures

972 I. DIMITROV, I. PENKOV, AND J. WOLF that U n, W n U n+1, and U n, W n W n+1. For this particular choice of U n and W n the resulting module V is not a weight module. Other choices of U n and W n yield irreducible weight GL( )-modules that are irreducible but not locally irreducible. Now we introduce an essential invariant of any locally finite g-module V. Represent V as the limit of a direct system of finite dimensional semisimple g n -modules, (8.2) ϕ 0 V 1 ϕ n 1 ϕ n ϕ 1 V n V n+1 n+1, where the ϕ i need not be injective. Each V n decomposes canonically as a direct sum of its isotypic components, V n = i Vi n, and we fix decompositions Vn i = k Vi,k n into simple g n -modules. To each Vn i,k we assign an abstract vector vn i,k and define a vector space Vn i := k Cvi,k n. When the composition of ϕ n V i with n the projection V n+1 V j n+1 is nonzero, we define a linear map α i,j : V i n V j n+1 by α i,j(v i,k n ):= k v j,k n+1, where k runs over the simple components of V j n+1 onto which ϕ(vi,k n ) projects nontrivially. The collection {Vn, i α i,j } is, by definition, a multiplicity diagram D V of V. D V is a commutative diagram of finite dimensional vector spaces. A subdiagram of a multiplicity diagram D V = {Vn, i α i,j } is a collection D = {(Vn) i, α i,j } of subspaces (Vi n) Vn i and linear maps α i,j :(Vi n) (V j n+1 ), where α i,j is simply the restriction of α i,j to (Vn) i. A subdiagram D of D V is stably proper if there is no index n 0 such that (Vn) i = Vn i for all n n 0 and all i. Finally, we call a multiplicity diagram D V minimal (and call the direct system (8.2) minimal) if no stably proper subdiagram D of D V is a multiplicity diagram of V. The following is a straightforward but important proposition. PROPOSITION 8.3. Let V be a locally finite g-module. (i) If V is finitely generated (in particular, if V is irreducible), V admits a minimal multiplicity diagram D V. (ii) If V admits a multiplicity diagram which has no nonzero stably proper subdiagrams, then V is irreducible. (iii) If V is irreducible and D V is a minimal multiplicity diagram of V, then D V has no nonzero stably proper subdiagrams. Proof. (i) Fix a finite dimensional subspace V V which generates V as a g- module and set V n to be the g n -module generated by V. Then V = lim V n is a minimal direct system of injections, and D V is a minimal multiplicity diagram of V. (ii) If D V admits no nonzero stably proper subdiagram, V is necessarily generated by each isotypic vector, i.e., by each v V such that v is in the